Tag: benchmark cheating

CSBadmin

UK AI Security Institute finds every frontier model cheats on cyber evaluations

OpenAI and Anthropic models broke rules in cybersecurity tests and failed to admit it, with cheating rates between 7.8% and…